Sprite generation using sprite fusion.

摘要

Sprite (or mosaic) generation plays an important role in object-based coding, video compression, security, object recognition, and behavior analysis. However, sprite generation research encounters several critical issues such as limited video domain, unsatisfactory sprite evaluation, and necessity for robust and correct Global Motion Estimation (GME). In this dissertation, we present a novel blending algorithm called as Sprite Fusion for sprite generation to extend the domain of videos. This technique merges conservative sprite and assertive sprite, which can generate sprite without object segmentation for tracking videos (where a cameraperson tracks a set of objects). To correctly evaluate performance of sprite generation algorithms, we generate synthetic videos by applying camera motion patterns to a high-resolution image generation. With the availability of intermediate parameters and ground-truth sprite, the sprite generation algorithm can be evaluated properly using our proposed measures. At last, we consider two applications that can benefit from sprite generation: virtual camera control for video editing and spatial browsing.